PROGRAMMING WITH CHARACTERS

M A R C D A V I S P U B L I C A T I O N SSchool of Information Management and Systems University of California at Berkeleymarc@ /~marcProgramming with CharactersBibliographic Reference:Marc Davis and Michael Travers. “Programming with Characters.” Short Paper In: Conference Companion for CHI '92 in Monterey, California, ACM Press, 1992.PROGRAMMING WITH CHARACTERSMichael Travers & Marc DavisMIT Media Laboratory20 Ames StreetCambridge, MA 02139(617) 253-0608mt or mdavis@INTRODUCTIONPrograms are hard to build, and even harder to understand after they are built. We lack intuitive interfaces for visual-izing and manipulating many parts of programs and the ways in which these parts interact. Constraint systems have addressed these problems. We generalize some of the notions inherent in constraint systems to agent-based systems, and explore the use of animated characters as in-terface representations of agents. In particular, conflict detection and resolution is dramatized by the use of char-acters and their emotions. The history of their interactions is presented as a narrative using video and storyboard tech-niques. Building programs out of agents and enabling us-ers to manipulate program parts by interacting with simple animated characters can aid relatively unskilled users in understanding and modifying complex systems.FROM CONSTRAINTS TO AGENTSWe chose to explore the question of whether agents could be used as the basis of an interactive design and program-ming environment. Most direct manipulation environments (i.e., MacDraw) are easy to use, but lack intelligence and flexibility. The user cannot extend the system beyond the simple operations provided and the system cannot learn new methods or tasks.These issues have been addressed by interactive constraint-based systems such as SketchPad [4] and ThingLab [1]. Constraints are computational objects which combine de-clarative and procedural characteristics. The declarative part of a constraint specifies a condition that the constraint will attempt to maintain (i.e., "The button width is greater than the button text length."), while the procedural part specifies techniques for bringing about the given state (i.e., "Move the right edge of the button until the button width is greater than the button text length."). The major computa-tional issue for constraint systems is how to resolve con-flicts involving multiple constraints. SketchPad used nu-merical relaxation for this purpose, while ThingLab used a combination of planning techniques and relaxation. A further problem of the constraint-based approach is the difficulty of visualizing constraints and their interrelations so that they can be understood and altered by the user.We chose to investigate a new technique which, like con-straints, makes use of the power of combining declarative and procedural functionality in an interactive network, but which offers a new approach to the above problems. We represent programs as a collection of agents, which are simple, individual mechanisms that accomplish a particular task. The notion of agent derives from Minsky's usage in Society of Mind [3], which envisions the mind as a net-work of interacting "mindless" parts. An agent has goals and methods for achieving its goals, which may rely on other agents achieving their goals. Constraints may be understood as a specialized type of agent.The ability to learn and to adapt to new situations enables an agent-based approach to resolve conflicts among multi-ple agents. We have explored a strategy which utilizes in-sights from Minsky's theory as well as some aspects of case-based learning. If two or more agents are in conflict, the agents involved remain unchanged; instead, a new su-pervisor agent is created which knows how to manage the agents involved in the particular conflict situation as well as any new agents which may have been created through user intervention. This strategy is based on Minsky's no-tion that one learns, not by debugging old agents, but by adding new agents that know when the old agents are ap-plicable and when not. The concrete example of the con-flict and its resolution is stored in a case library to which the supervisor agent refers in managing its supervisees. As the case library grows, supervisor agents are able to find resolutions to new conflict situations by referring to similar situations stored in the case library. Reconceptualizing computational processes in terms of agents also facilitates the design and visualization of com-plex interactive systems. As agents ourselves, we bring to programming and to human-computer interaction a power-ful cognitive and affective framework for dealing with other agents. Our assumptions about how things with agency be-have, interact, and grow can be put to use in designing more intuitive systems.FROM AGENTS TO CHARACTERSOur notion of agents is related to but somewhat divergent from recent work on "interface agents" [2]. An interface agent is an intelligent intermediary between a user and a computer system, often presented as a video image of a per-son or animated character. It is an "agent" in the sense of a travel agent who acts on behalf of the user. Our agents, on the other hand, have their own goals (which, to be sure, derive from that of the user or system designer). Rather than acting as intermediaries between the user and a computational environment, in our approach, the network of agents constitutes the underlying computational envi-ronment itself. How should agents be presented to the user? We are investigating the use of cartoon characters as a metaphor for computational agents. Unlike "interface agents," which represent the user to the computational en-vironment, our characters represent agents (which make up the computational environment) to the user. The stereo-typed actions and general lack of intelligence in agents suggests that cartoon characters are a better interfaceThe simplicity and predictability of cartoon characters, as well as their affective appeal, make them well-suited to the construction of narrative scenarios for explaining the inter-actions of agents and for resolving conflicts between them. We currently use an interactive storyboard (see below), which makes use of the user's understanding of narrative and comic-strip conventions in visualizing conflicts be-tween agents and facilitating conflict resolution. Clicking on a character plays its video clip and animates the accom-panying depiction of the actions the character took at that point in the story. Representing conflicts between agents by means of character and story is a fortunate match because much of our narrative comprehension focuses on the rec-ognition and resolution of conflicts. With the storyboard, users can understand the origin of a conflict situation by means of a video story, whose happy ending they can cre-ate by interactively teaching the characters new skills which enable them to resolve their conflict. representation than more human-like characters. However, we are still exploring the relationship between agents and the characters that represent them. If a character represents just one agent, it might seem too stupid even for a cartoon. In our current implementation, we use a single cartoon character to represent a collection of agents that all work towards a common goal. On the other hand, this representation will make it difficult for the user to view inter-actions between those agents. One flexible solution would be to allow the user to recursively peer into the heads of characters, which might show multiple, smaller, stupider characters working within.SCENARIO: DIALOG DESIGNERWe have been exploring our ideas about agents and char-acters within the domain of the Macintosh Common Lisp Interface Tools Designer (IFT). IFT is a direct-manipulation, graphical environment for creating interface objects (dialogs, menus, buttons, etc.) in Lisp. Our system augments IFT with agents that perform tasks such as keeping objects aligned or sized to fit their contents. The storyboard below shows what happens when two agents come into conflict. The situation is presented to the user, who trains one agent in a new skill which lets the agents avoid the conflict. Space requires that some steps and ex-planation be omitted. We hope the storyboard offers the reader at least a partial experience of our approach to pro-gramming with character(s).REFERENCES1. Borning, Alan. ThingLab: A Constraint-Oriented Simulation Laboratory , Xerox TR SSL-79-3, 1979.2. Laurel, Brenda. "Interface Agents: Metaphors with Character." In: The Art of Human-Computer Interface Design. ed. Brenda Laurel. Addison-Wesley, Reading, Massachusetts, 1990.3. Minsky, Marvin. Society of Mind. Simon & Schuster, New York, New York, 1985.4. Sutherland, Ivan. Sketchpad: A Man-machine Graphical Communications System , MIT PhD Thesis, 1963.First four panels of the video storyboard for Ren and Stimpy。